DeepSeek builds high-performing open-weight models that rival closed frontier labs on reasoning and coding while costing dramatically less to run. The open licenses make it a favorite for teams that want frontier quality they can self-host or fine-tune.
LLMs & AssistantsOpen SourceEmergingFrom Free (open weights) / API from $0.15 per M input tokens off-peak (V4.1-Flash)
＋Frontier-class reasoning and coding at a small fraction of the cost
＋Open weights allow self-hosting, fine-tuning and full data control
＋API pricing undercuts most closed competitors substantially
Limitations
－Hosted service raises data-residency questions for regulated buyers
－Enterprise support and certifications lag Western labs
－Guardrails and content policies differ from US-based providers
Evidence behind this score
Each score is built from the claims below. This is a snapshot, so check the current terms with the vendor before you commit to anything.
2 of 5 claims below are past their re-check window. Pricing, availability and security details move fast, so confirm those with the vendor.
Claim
Source
Checked
Open-weight releases matched leading models on reasoning benchmarks at far lower cost· re-check
Public leaderboards & DeepSeek releases
Feb 2026
Weights distributed under permissive licenses for self-hosting
DeepSeek model cards
Dec 2025
API priced well below comparable closed models· re-check
DeepSeek pricing page
Mar 2026
V4.1 Flash (Sept 10, 2026) replaced V4-Flash pricing with MIT-licensed weights and off-peak rates of $0.15/M input, $0.60/M output (peak: $0.30/$1.20, 01:00-04:00 and 06:00-10:00 UTC weekdays) — half the prior Flash tier; it also absorbs all V4-Pro traffic at Flash pricing from Sept 14 until a dedicated V4.1-Pro ships
A joint NSA/CISA/FBI advisory (AA26-251A, Sept 8, 2026) named DeepSeek among six China-based firms it says ran an organized distillation campaign against US frontier models since late 2024, extracting outputs from Claude, GPT and Gemini variants to train R1 and V3; the advisory recommends providers quietly degrade rather than block flagged accounts
Anthropic's threat report: Russian hackers used Claude to auto-tune malware past antivirus
Anthropic's September threat intelligence report says the Russian state-linked group Midnight Blizzard used Claude to check whether its malware evaded detection by security products, then had AI agents automatically patch and rebuild any sample that got flagged and redeploy it, repeating the loop until it slipped through. More than 20 organizations were targeted. Separately, a Russian-speaking criminal actor combined OpenAI and DeepSeek models to run hundreds of AI agents against a zero-day pair in PaperCut NG/MF, compromising at least 440 servers across 395 organizations in 48 countries before a patch landed. Neither campaign needed a novel exploit; both needed an agent that could iterate faster than a human defender could watch.
DeepSeek undercuts its own Flash tier again, down to $0.15 per million input tokens
DeepSeek released V4.1 Flash with MIT-licensed weights, a 552B-parameter mixture-of-experts design on a new causal encoder-decoder architecture, and off-peak pricing of $0.15 per million input tokens and $0.60 output — half of what V4 Flash charged, with peak-hour rates of $0.30 and $1.20 for anyone who needs the model outside its cheapest hours. Cache hits drop to $0.003 off-peak. The model takes over V4 Pro's traffic entirely on September 14 until a dedicated V4.1 Pro ships, so anyone still calling the Pro endpoint gets routed to Flash-tier weights at Flash-tier prices without changing a line of code.
NSA, CISA and FBI accuse six Chinese AI firms of systematic model distillation
A joint advisory (AA26-251A) from the NSA, CISA and FBI says DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI ran an industrial-scale campaign since late 2024 to extract outputs from Claude, GPT, Gemini and Grok and use them as synthetic training data. The advisory says this isn't a side channel but the core of how those firms built models like DeepSeek's R1 and V3, and it recommends US providers quietly degrade flagged accounts rather than block them outright, so the distillation doesn't simply move to a new account. It's the first time the agencies have named specific companies rather than describing the technique in the abstract.
DeepSeek's price hike lands: peak-hour V4 rates up to 12x the old flat rate
The 'significant increase' DeepSeek flagged on August 6 took effect August 16 at 16:00 UTC, replacing flat per-token billing with peak/off-peak pricing (peak: 01:00-04:00 and 06:00-10:00 UTC, off-peak at half that). V4-Flash output jumps from a flat $0.28/M to $1.32/M at peak; V4-Pro output rises from $0.87/M to $3.96/M at peak — increases of 50% to over 1,000% depending on tier, with the cache-hit tier hit hardest. Even at peak rates DeepSeek stays well under frontier closed-model pricing. But the era of treating it as a free floor for volume work is over, and anything scheduled during UTC peak hours needs pricing again.
DeepSeekDeepSeek API changelog and pricing page; Bloomberg, Pandaily and Engadget coverage